在FPGA平台利用深度學習進行智慧監控

Intelligent Surveillance Using Deep Learning on FPGA Platform

本文旨在針對運算資源和功率限制下，尤其是在嵌入式邊緣設備上進行即時推理的挑戰，如何有效地優化訓練和推理速度成為當前熱門的研究課題。YOLO（You Only Look Once）演算法因其只需對影像進行一次卷積神經網絡（CNN）處理，便能實現高效能和高偵測率，相比於兩階段的物件辨識工具（如R-CNN、Faster R-CNN等），YOLO演算法顯著提升了辨識速度。
在此背景下，本研究致力於在FPGA-KV260平台上結合深度學習技術，實現智慧監視系統。KV260作為一款高效能的FPGA平台，能夠充分發揮深度學習演算法的潛力，提升運算效率並節省功耗。本次實驗旨在使用YOLOv4演算法，達成以下目標：
l生成YOLOv4權重檔（Weights File）：通過YOLOv4模型的訓練，結合適當的訓練集，生成專案所需的權重檔。
l串流影像至KV260並進行即時偵測：利用網路攝像頭（Webcam）將影像資訊串流至FPGA-KV260平台。透過KV260上的YOLOv4模型進行人臉偵測，並將偵測結果在電腦端顯示。實驗計畫表明，所提出的實現方法在物件光線等外在條件充足的情況下，可以達95%以上的準確率。

This paper focuses on addressing the challenges of real-time inference under computational resource and power limitations, especially on embedded edge devices. Optimizing both training and inference speeds efficiently in these contexts has become a significant research topic. The YOLO (You Only Look Once) algorithm, which processes images using a single pass through a convolutional neural network (CNN), achieves high performance and detection rates. Compared to two-stage object detection models like R-CNN and Faster R-CNN, YOLO significantly increases recognition speed.
In this context, the research aims to develop an intelligent surveillance system by integrating deep learning techniques on the FPGA-KV260 platform. The KV260, as a high-performance FPGA platform, maximizes the potential of deep learning algorithms, enhancing computational efficiency and reducing power consumption. This experiment aims to achieve the following objectives using the YOLOv4 algorithm：
lGenerate YOLOv4 Weights File：By training the YOLOv4 model with an appropriate dataset, a weights file required for the project will be generated. Creating a custom training dataset ensures higher accuracy compared to open-source options and aligns anchor boxes with project-specific needs, minimizing unwanted detection results.
lStream Video to KV260 for Real-Time Detection：A webcam will stream video data to the FPGA-KV260 platform, where the YOLOv4 model will perform facial detection, with results displayed on a computer. The experimental results indicate that the proposed method can achieve over 95% accuracy under favorable lighting conditions.